Supervised Learning From Embedded Subgraphs

Abstract

We develop a machine learning algorithm which learns rules for classification from training examples in a graph representation. However, unlike most other such algorithms which use one graph for each example, ours allows all of the training examples to be in a single, connected graph. We employ the Minimum Description Length principle to produce a novel performance metric for judging the value of a learned classification. We implement the algorithm by extending the Subdue graph-based learning system. Finally, we demonstrate the use of the new system in two different domains, earth science and homeland security.